From Constraints to Finite Automata to Filtering Algorithms

From Constraintsto Finite Automata

to Filtering Algorithms

Mats Carlsson, SICSNicolas Beldiceanu, EMN

[email protected]

[email protected]

ESOP, March 29, 2004

Outline Constraint Propagation: Example &

Model Constraints and Key Notions Case Study: X lex Y

Definition and signature Finite automaton Filtering algorithm

Case Study: lex_chain(X0,…,Xm-1) Definition and signature Finite automaton Filtering algorithm

Conclusion


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Example

x + y = 9 2x + 4y = 24

x 0 1 2 3 4 5 6 7 8 9

y 0 1 2 3 4 5 6 7 8 9


Constraint Propagation Variables

feature variable domain (finite set of integers)

Propagators implement constraints

Propagation loop execute propagators until simultaneous

fixpoint


Propagator Propagator p is a procedure

(coroutine) implements constraint con(p)

its semantics (set of tuples) computes on set of variables var(p)

Execution of propagator p filters domains of variables in var(p) signals failure signals entailment


Propagators Are Intensional Propagators implement filtering

aka: narrowing, domain reduction, value removal

No extensional representation of con(p) impractical in most cases (space)

Extensional representation of constraint can be provided by special propagator often: “element” constraint, “relation” constraint,

…


Propagation Events

Normally, a propagator p is resumed whenever some value in a domain of var(p) has been removed.

In some cases, some events (e.g. removing internal values) are irrelevant whilst other (bounds adjustments) are relevant.


Idempotent Propagators A propagator is idempotent if it

always computes a fixpoint.

Most constraint programming systems can accommodate both idempotent and non-idempotent propagators.


Implementing Propagators Implementation uses operations on

variables reading domain information filtering domains (removing values)

Variables are the only communication channels between propagators

Algorithms for Domain filtering Failure detection Entailment detection






Conclusion


Classes of Constraints Basic constraints

Constraints for which the solver is complete x D, x = v, x = y (variable aliasing)

Primitive constraints (need propagators) Non-decomposable constraints

x<y, xy, x+y = z, x*y = z, …

Global constraints (need propagators) Subsume a set of basic or primitive

constraints, usually providing stronger consistency


Support and Consistency Given: constraint C, variable x var(C),

its domain D(x), integer v. x=v has support for C iff

v D(x) C has a solution such that x=v

C is hyperarc consistent iff x var(C) v D(x) x=v has support for C

Maintaining hyperarc consistency may not be possible with polynomial algorithms (e.g. diophantine equations)


Entailment A constraint con(p) is entailed if it

holds for any combination of values in the current domains.

Consequences for its propagator p: It has no more work to do It should not be resumed any more (up to

backtracking) It is usually reponsible for detecting

entailment


Failure A constraint con(p) is false if it does

not hold for any combination of values in the current domains.

Consequences for its propagator p: It should signal inconsistency, e.g. by

instigating backtracking It is reponsible for detecting failure


Notation

Vectors and subvectors X = (x0,…,xn-1) X[0,r) = (x0,…,xr-1), r n

Domain variables D(x), the domain of x (set of integers) min(x), lower bound of x, O(1) max(x), upper bound of x, O(1) prev(x,b) = max{y D(x) | y<b}, O(d) next(x,b) = min{y D(x) | y>b}, O(d)


Constraint Signatures The constraint store is the set of all

domains D(x) For alphabet A, constraint C,

constraint store G, let S(C,G,A) be the signature of C wrt. G and A.

The filtering algorithm is derived from a finite automaton for signatures.






Conclusion


Definition:X lex Y Let:

X = (x0,…,xn-1) Y = (y0,…,yn-1) xi, yi domain variables or integers

X lex Y holds iff n=0, or x0<y0, or x0=y0 (x1,…,xn-1) lex(y1,…,yn-1).


Signature: X lex Y

Letter

Condition

< max(xi)<min(yi)

= xi = yi, integers

> min(xi)>max(yi)

max(xi)=min(yi)min(xi)<max(yi)

min(xi)=max(yi)max(xi)>min(yi)

? otherwise

$ End of string


Signature example: X lex Y

X 3..3

3..4

4..5

4..5

3..3

4..4

Y 3..3

4..5

3..4

4..5

4..4

3..3

S = ? < > $


Poset of signature letters

< = >

?

E.g., a becomes a < or a = in a ground store.






Conclusion


Finite Automaton for X lex Y

1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$

$ $ $


Success State T1

1

T1

$

Enforce xi=yi in the leading prefix for C to hold.Afterwards, the leading prefix is ground and equal.


Success State T2

1 2 4

T2

<q

Enforce xq<yq in order for there to be at least one < preceding the first >.


Success State T3

1 2 3

T3

$ $

q

Only enforce xqyq , for < can appear in a later position.


Delay States

1 2 3 4

T3 T2

D1 D3 D2

$

$ $

q

Not yet enough information to know what to do at position q.

T1

$






Conclusion


Filtering Algorithms Non-incremental, O(n)

Run finite automaton from scratch Consider all letters from scratch

Incremental, amortized O(1) Deal with one letter change at a time Needs to know what letter has

changed, in what state


Incremental Restart 1

1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$

$ $ $

Resume in state 1.



1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$

$ $ $

Resume in state 2.



1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$

$ $ $

Resume in state 3 or 4, resp.



1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$

$ $ $

If changed to =, no-op. Otherwise, resume in state 3 or 4, resp.


Finite Automaton for X <lex Y

1 2 3 4

T1 T3 T2

F1 D1 D3 D2

$ $ $ $






Conclusion


Definition:lex_chain(X0,…,Xm-1)

Let: Xi = (xi0,…,xin-1) xij domain variables or integers

lex_chain(X0,…,Xm-1) holds iff

X0 lex … lex Xm-1


Internal constraint:between(A,X,B) Preconditions:

A = (a0,…,an-1), B = (b0,…,bn-1) X = (x0,…,xn-1) ai,bi integers; xi domain variables i[0,n) : ai D(xi), bi D(xi)

Holds iff:A lex X lex B


Signature:between(A,X,B)

Letter

Condition

< ai<bi next(xi,ai)bi

« ai<bi next(xi,ai)<bi

= min(xi)=ai=bi=max(xi)

# min(xi)ai=bi max(xi)ai=bi

> ai>bi bimin(xi)max(xi)ai

» ai>bi (min(xi)<biai<max(xi))

$ End of string


Signature example: between(A,X,B)

A 5 4 6 6

X 4..6 4..6 4..6 3..7

B 5 5 4 4

S # < > » $

X’ 5..5 4..5 {4,6}

{3,4,6,7}






Conclusion


Finite Automaton:between(A,X,B)

1 2

T1 T2

F1

=#

>»

«$ <«»#$

>=

<

State 1 denotes a prefix in which ai=bi. Hence we must enforce xi=ai=bi there.


Success State T1:between(A,X,B)

1

T1

=#

«$Either q=n or xq=v has support for all aqvbq.Hence we enforce aqxqbq.

q


Success State T2:between(A,X,B)

1 2

T2

=#

<«»#$

>=

<

We have:X[0,r)=A[0,r)X[0,r)=B[0,r)

Hence we enforce:aixibi for i[0,r).

Either r=n or xr=v has support for all vbr var.

Hence we enforce:xrbr xrar.

r






Conclusion


Feasible Upper Bound:Problem Given X and B, compute the

lexicographically largest U such that:

U lex B i[0,n) : ui D(xi)

Similarly for feasible lowest bound


Feasible Upper Bound:Algorithm

Compute as the smallest i such thatU[0,i)B[0,i)

Compute: ui = bi, if i < ui = prev(xi,bi), if i =

ui = max(xi), if i > Similarly for feasible lowest bound


Filtering:lex_chain(X0,…,Xm-1)

1. Compute feasible upper bound Bi for Xi from i=m-1 down to i=0.

2. Compute feasible lower bound Ai for Xi from i=0 to i=m-1.

3. Enforce Ai lex Xi lex Bi for all i.

Arc-consistency in O(nmd) time.






Conclusion


Results An approach to designing filtering

algorithms by derivation from FAs on constraint signatures

Case studies and hyperarc consistency algorithms for two constraints: X lex Y, running in amortized O(1) time per

propagation event lex_chain(X0,…,Xm-1), running in O(nmd)

time per invocation


Future Work What constraints are amenable to the

approach? Where does the alphabet come from? Where does the automaton come from? Where do the pruning rules come from? How do we make the algorithms

incremental?


References and proofs SICS T2002-17: Revisiting the

Lexicographic Ordering Constraint, Mats Carlsson, Nicolas Beldiceanu.

SICS T2002-18: Arc-Consistency for a Chain of Lexicographic Ordering Constraints, Mats Carlsson, Nicolas Beldiceanu.

http://www.sics.se/libindex.html


Related Work Global constraints for

lexicographic orderings. A. Frisch, B. Hnich, Z. Kızıltan, I. Miguel, T. Walsh. Proc. CP’2002. LNCS 2470, pp. 93-108, Springer, 2002.

Documents

From Constraints to Finite Automata to Filtering Algorithms